|
Abstract:
|
Magnetic resonance imaging is crucial for detection and characterization of structural changes in the brain associated with neurological and psychiatric disease. Such phenotypes use MRI technology and segmentation approaches, manual and automatic, for delineating structures of interest in the brain. Numerous automatic approaches result in location-specific maps of the probability of a structure of interest lying in each voxel, and a threshold is then applied to generate binary segmentation masks. Automatic approaches often involve an expert manually defining a threshold for abnormality which is used across the image to optimize sensitivity and specificity. However, this does not allow for differential thresholding by location which can result in improved global segmentation performance. We propose a general dynamic approach, ADAPT, for data-driven spatially adaptive thresholding. The method utilizes healthy controls to generate an empirical null distribution that any voxel does not belong to the tissue class of interest. We believe the proposed methods will easily extend to thresholding of probability maps in many segmentation and abnormality detection image analysis settings.
|
